Decentralized POI Recommender Systems

The School of EECS is hosting the following PhD Progress Review 3 Thesis Review Seminar:

Decentralized POI Recommender Systems

Speaker: Jing Long
Chair: A/Prof. Guangdong Bai

Abstract: As an indispensable personalized service in Location-based Social Networks (LBSNs), the next Point-of-Interest (POI) recommendation aims to help people discover attractive and interesting places. Traditional cloud-based POI recommenders, while effective, face limitations such as privacy concerns and low robustness. To this end, on-device and decentralized models have gained attention for their ability to provide personalized recommendations while safeguarding user data. In detail, the thesis proposes several key works as solutions in decentralized POI recommender systems:

Diffusion-Based Cloud-Edge-Device Collaborative Learning (DCPR) for Next POI Recommendations: This model leverages a multi-layer architecture involving the cloud, edge servers, and user devices. It enables the deployment of lightweight models on devices for local fine-tuning and inference, adapting quickly to new regions and users, while maintaining efficiency in memory and time usage.

Decentralized Collaborative Learning Framework (DCLR): DCLR enhances on-device recommendations by utilizing self-supervised learning to capture geographical and categorical correlations. It identifies neighbors based on semantic and geographical similarities, ensuring effective knowledge transfer across devices and improving recommendation accuracy in sparse data environments.

Model-Agnostic Decentralized Collaborative Learning (MAC): This framework allows users to customize their model configurations and store embeddings for POIs relevant to their context. It incorporates performance-triggered and similarity-based sampling techniques to optimize resource use and enhance both recommendation quality and privacy.

Physical Trajectory Inference Attack (PTIA) and Adversarial Defense Mechanism (AGD): The review also explores privacy vulnerabilities in decentralized learning systems. PTIA is introduced as an attack that can infer users' mobility patterns. To counter this, the adversarial defense mechanism (AGD) effectively obscures sensitive data, ensuring robust privacy protection without sacrificing recommendation accuracy.

The experimental results validate the superiority of these proposed methods over existing models, highlighting significant improvements in recommendation accuracy, memory efficiency, and privacy protection. This review sets the stage for future research into more secure and effective decentralized POI recommender systems.

Biography: Jing Long is a dedicated PhD candidate at the University of Queensland, specializing in decentralized recommendation systems. With a strong foundation in data science and machine learning, Jing has demonstrated expertise in enhancing user privacy and system efficiency in recommendation solutions through innovative research. Currently pursuing a Doctor of Philosophy in Computer Science, Jing has developed a novel framework for decentralized Point-of-Interest (POI) recommender systems, addressing critical challenges such as privacy protection and resource consumption. Their research has resulted in four publications as the lead author in CCF A-ranked conferences and journals, showcasing their commitment to advancing the field of artificial intelligence. Jing's work has been presented at three international conferences, where they received positive feedback for its innovative nature and practical applicability. In addition to research, Jing has valuable experience as an academic tutor, facilitating discussions and fostering interactive learning environments in courses related to data analytics and social media. Their practical experience includes a role as a data scientist at Rio Tinto, where they developed custom analytics tools and collaborated closely with teams to optimize operations through data insights.